Show HN: ZQK โ a microkernel OS for autonomous agents (local-first, capability-secured)
Tagline: Stop building agents on prompt wrappers. ZQK is a host-level state machine that treats the LLM as an unprivileged inference device.
The problem
Every production AI agent hits the same wall:
- Non-determinism: The LLM decides what state to enter, what tool to call, and where to route next. When it hallucinates, your agent enters undefined behavior.
- Memory bloat: Agent "state" lives in the prompt. After 20+ turns, you're spending thousands of tokens reconstructing context that should be in a database.
- No real security: "Capability" means a system prompt saying "don't delete things." The LLM can be jailbroken. There's no kernel enforcing anything.
- No multi-agent coordination: Two agents sharing a task means coordinating via prompt injection. Race conditions and overwrite conflicts are invisible.
What ZQK does differently
Traditional: LLM โ decides next tool โ decides next state โ hopes for the best
ZQK: Host FSM โ calls LLM as a plugin โ receives structured output โ drives next state
The kernel, not the model, is in control.
Deterministic host FSM
State transitions live in Go code (pkg/pipeline/orchestrator.go). The LLM returns structured data โ the FSM decides what happens next. No next_state from model strings. Ever.
Capability-secured operations
Every storage operation goes through a role enforcement gate. Admin operations require an explicit key. Agents can't escalate privilege via a clever prompt.
Kernel-managed state (not prompt state)
Task state lives in a structured graph store (Neo4j/MemGraph) โ not in the conversation. zqk workflow whats-next always gives the agent accurate current state in ~50 tokens, regardless of session length.
Native MCP (JSON-RPC 2.0)
The community binary exposes tools/list and tools/call over stdio and TCP. Works with Cursor, Claude, VS Code, and any MCP client out of the box.
Apache / Enterprise delineation ACK
| Tier | License | What you get |
|---|---|---|
| Community kernel | Apache License 2.0 (LICENSE / NOTICE) |
Single-node cell: CLI, MCP, scheduler, local graph/CAS |
| Enterprise | Separately licensed (commercial) | Multi-node fleet mesh, RBAC, cross-cell compliance, sealed fleet audit |
Community Apache โ Enterprise entitlement. Evaluating the open-core tree does not grant Enterprise modules.
Quick start (< 2 min)
Aligned with Community First-Run and 60-second capability demo:
# Install (public binary โ go install โ source; no GITHUB_TOKEN required)
curl -sSL https://raw.githubusercontent.com/zqk-os/zqk/main/scripts/install.sh | sh
# First agent:
mkdir my-project && cd my-project
zqk init
zqk workflow whats-next # kernel tells agent current mission
zqk mcp ensure # wire Cursor / Claude Code
Buyer smoke (~60s): see docs/demos/SIXTY_SECOND_CAPABILITY.md.
Comparison
| LangChain / CrewAI | AutoGen | ZQK | |
|---|---|---|---|
| State lives in | Prompt text | Prompt text | Kernel (graph store) |
| State machine | LLM decides | LLM decides | Host FSM (Go code) |
| Capability enforcement | System prompt | System prompt | Kernel gate |
| Multi-agent coordination | Prompt injection | Message passing | Shared kernel (atomic CAS) |
| Init time | 38โ52s | 52s+ | ~180ms |
| Memory (idle agent) | 95โ120MB | 120MB+ | ~18MB |
| MCP support | Plugin | Plugin | Native (first-class) |
| Language | Python | Python | Go (precompiled binary) |
Who this is for
- Infra/backend engineers burned by non-deterministic agent loops in production
- Local-first / privacy teams running Ollama, Apple Silicon, local GPU (no cloud required)
- Platform teams building multi-agent systems that need real coordination, not prompt negotiation
What's in the box
zqk system initโ scaffold structured project (goals, plans, backlog, milestones)zqk workflow whats-nextโ kernel-driven mission discovery (no prompt reconstruction)zqk schedulerโ background task runner with retry, concurrency limits, test bundlingzqk mcp serve --stdioโ MCP JSON-RPC 2.0 server (Cursor/Claude/VS Code)zqk objectโ full CRUD for all kernel object typespkg/pipeline/orchestrator.goโ embeddable host FSM for your own agent workflows
Open-core model
| Tier | License | Contents |
|---|---|---|
| Community kernel | Apache License 2.0 | Single-node kernel, MCP, scheduler, local graph/CAS |
| Enterprise | Separately licensed (commercial) | Multi-node fleet mesh, enterprise IdP/RBAC, sealed fleet audit |
See the ACK table above. No MIT community SKU remains on this trust surface.
Links
- Architecture โ host FSM ยท capability flow ยท paging
- Guides & Recipes โ caged worker ยท MCP DB agent ยท long-running research
- 60-second capability demo
- Quickstart
- Community First-Run
- Security Primitives & Sanitation
Built by the ZQK team. Local-first, deterministic, and production-ready for AI + human hybrid teams.